Operator request: give the webbug some personality. Trigger response is now
a small JPEG embedded into the binary via //go:embed, served from
backend/internal/token/generators/webbug/asset/pixel.jpg. Default is a
545-byte 32x32 yellow placeholder I generated with imagemagick — replace
the file with any image (same path, same filename, doesn't matter what
size or content) and the next backend rebuild picks it up. No other code
needs to change.
Scope: ONLY webbug. The docx, pdf, and envfile generators keep their own
Trigger functions that still call pixel.Clone() for the canonical
invisible 1x1 GIF — those types embed the trigger URL inside their
respective artifact bytes (a Word footer, a PDF /AA action, a fake .env
INTERNAL_METRICS_ENDPOINT line) and need the response to stay invisible
so the attacker doesn't notice they pinged us. Stealth preserved where
it matters.
The pixel package itself is untouched (still has the 43-byte GIF
constant); only webbug stops importing it.
Tradeoff baked in: a webbug whose response is a visible image is
*technically* less stealthy than the canonical 1x1 — the attacker can
see "huh there's a small image here." But the visible mechanism is real
too (think Mailchimp / SendGrid email-open tracking, which use sized
"social proof" images, not just 1x1s), so this is a legitimate variant.
For an operator who wants strict 1x1-invisible behavior in prod, the
revert is a 3-line change in webbug/generator.go (swap embed for
pixel.Clone() again).
Tests updated: webbug/generator_test.go no longer asserts gifByteLength
or pixel.Clone() equality; it now asserts the body starts with the JPEG
SOI marker (0xff 0xd8) + ContentType is image/jpeg. The
"independent-copy-per-call" invariant test still passes (bytes.Clone in
the Trigger ensures each Body slice is a fresh allocation).
Cache headers unchanged (no-store / no-cache / must-revalidate) so every
fetch hits the trigger handler.
Upstream 8d2599c0 added a mysqlEnabled parameter to token.NewHandler and
took a long line over the limit in handler_test. Pass the parameter from
the e2e stack (true: integration test exercises mysql token creation) and
reflow the long handler_test return so golines is happy.
Operator request: when MYSQL_FAKE_ENABLED=false (the dev default + the only
viable setting behind Cloudflare Tunnel since CF Tunnel doesn't carry raw
TCP), surface that in the UI so users can see the species but understand
they can't deploy it on this instance. Avoids the silent-broken case where
someone creates a mysql token, gets a connection_string artifact, points
their MySQL client at it, and just times out with no clue why.
Backend (3-line change):
- TypeDescriptor gains `Enabled bool` + optional `DisabledReason string`
- TypeDescriptors() now takes `mysqlEnabled bool`; mysql is the one whose
Enabled tracks the flag, all others are always true (they don't need
any deployment-level infra beyond HTTP)
- Handler stores mysqlEnabled, passes to descriptors call; main.go wires
cfg.MySQL.Enabled in. Handler tests updated to pass false explicitly.
Frontend (matching schema + visual treatment):
- typeDescriptorSchema picks up `enabled` (default true for backward compat)
and `disabled_reason` (optional string)
- TypeCard reads descriptor.enabled, sets `data-disabled` on the label,
disables the radio input, replaces the blurb with the disabled_reason
when present (so the user sees WHY rather than guessing), and renders
a small "disabled" pill in the top-right corner of the card
- SCSS: dashed border, muted palette, opacity 0.55, glyph faded, cursor
not-allowed, hover effect suppressed — reads as "shelf is here but the
specimen is checked out" rather than a broken button
In the operator UI: mysql card renders ghosted with the explanatory text
"Requires direct TCP exposure (port 3306). Not reachable via Cloudflare
Tunnel — only enable on a VPS with raw TCP access." Selecting it is
mechanically impossible (input is disabled), no form-level guard needed.
Flip MYSQL_FAKE_ENABLED=true in env when you stand up a VPS later and the
card re-enables with no other code changes.
Backend builds clean + token-package tests pass.
Frontend gates green: typecheck + biome + lint:scss + build (431ms).
Audit F2: the PDF placeholder was padded with underscores embedded directly
in the /URI path, so Acrobat fetched /c/<id>____________ and the chi route
extracted the padded id. Lookup returned ErrNotFound, the webbug fallback
served a 1x1 GIF, and event recording skipped because tok was nil. PDF
tokens silently never recorded events.
Part B (durable): pad inside a ?p= query string so the trigger path stays
canonical and chi extracts the real id. Part A (defense in depth): handler
TrimRight on _ so PDFs already in the field also resolve correctly. Added
two regression tests: byte-after-id assertion in the PDF artifact, and the
handler's trim path.
- internal/admin/handler.go: replaces unused template stats handler
with the canonical Phase 12 endpoints from spec §8.1:
GET /stats — tokens_count, events_count, by_type,
by_alert_channel
GET /tokens — offset-based pagination, default 50,
cap 100; returns full token.Response
list with trigger/manage URLs via
injected URLBuilder; 400 BAD_PARAM on
negative/non-int offset
POST /tokens/{id}/disable — 204 on success; 404 NOT_FOUND envelope
when token.ErrNotFound; 500 envelope
on other repo errors
Handler takes TokenRepository + EventRepository + URLBuilder
interfaces (test seam); wire-up supplies *token.Repository,
*event.Repository, *token.Service (which already implements
TriggerURL/ManageURL).
- internal/admin/dto.go: Stats, TokenListPage, TokenListResponse.
- token.Repository.CountByType / CountByAlertChannel: GROUP BY type
and GROUP BY alert_channel; returned shapes typed as TypeCount /
ChannelCount with json+db tags for direct JSON emission.
- event.Repository.CountAll: global event count.
- Handler tests cover happy path + 500 propagation for each repo
dependency; pagination (default/limit-cap/offset paging); 400 on
bad offset; 404 on disable miss; 405 on GET to disable endpoint
(sanity check that POST-only routing is intentional).
Both audits flagged buildPage as a non-blocking but worth-cleaning nit:
it re-derived next_cursor from len(events) and the last event's ID
rather than using the HasMore + NextCursor flags that
event.Repository.ListByToken already computes via the LIMIT+1 peek
trick. Both calculations agreed today, but the duplication would
mask a divergence if the repo's pagination logic ever evolved
(e.g. switched to opaque token cursors).
Inline buildPage into gatherManageData and use list.HasMore +
list.NextCursor as the authoritative signals. Drops the buildPage
helper. Behavior identical; tests still pass.
Adds the HTTP layer for spec §8.4. The UUID in the URL is the
capability — no other auth required, by design.
- token/dto.go: ManageTokenView (slimmer than Response — no manage_id,
no manage_url, no metadata, since the user is already on the manage
page and those would be redundant), ManagePage{NextCursor, HasMore},
ManageResponse{Token, Events, EventsTotal, EventsSilencedActive, Page}.
Token.ToManageView(triggerURL) builder. Events use the existing
event.Response (already matches spec shape exactly).
- token/handler.go: EventQuery + DedupCounter interfaces declared near
the existing EventRecorder + FingerprintRecorder; same dependency-
injection style. NewHandler signature gains the two new args (5 → 6;
test sites + main.go updated).
- GetManage(w, r): parses ?cursor= (int64, must be >= 0) and ?limit=
(defaults 20, capped at 100). Returns 404 envelope for unknown
manage_id (envelope, not bare http.NotFound, so frontend gets a
parseable error code). 400 BAD_CURSOR for non-numeric or negative.
Calls svc.GetByManageID + eventQuery.ListByToken + CountByToken +
dedupCounter.CountActiveDedup; assembles the manage payload.
- DeleteManage(w, r): 204 on success, 404 on miss. Cascade-delete is
via the existing tokens-events FK ON DELETE CASCADE.
- buildPage helper: NextCursor is the string of the last event's ID
iff len(events) == limit (matches spec example "next_cursor": "41").
Tests cover happy path, 404 on unknown id, 400 on bad/negative cursor,
?limit= respected, ?limit=999 capped at 100, DELETE happy + 404.
Adds the two service primitives Phase 11's manage handler needs.
- event.Service.CountActiveDedup(ctx, tokenID): Redis SCAN over
dedup:trigger:{tokenID}:* in batches of 100, sums (value-1) for each
key (value=1 = first trigger fired notification, value=N>1 = N-1
silenced). Returns 0 when no rdb is configured.
- token.Service.DeleteByManageID: thin wrapper around
Repository.DeleteByManageID; ServiceRepository interface gains the
method so test fakes can implement it (fakeRepo updated). Returns
ErrNotFound on no-rows.
Tests cover: empty pattern returns 0, first-trigger-only returns 0
(notify fired, nothing silenced), aggregation across IPs sums correctly,
other tokens' keys are excluded by the prefix match, nil rdb returns 0.
Closes deferred Phase-9 item 8. Phase 9 promoted realIP / lastNonEmptyXFF /
OptionalHeader to internal/middleware and refactored webbug only; the
other generators kept their byte-identical local copies. Sweeping them
all over now while the touch surface is already open for Phase 10.
- docx, envfile, kubeconfig, pdf, slowredirect: drop the local
realIP / lastNonEmptyXFF / optionalHeader functions plus the
CF-Connecting-IP / X-Forwarded-For / X-Real-IP header constants.
Trigger paths now call middleware.RealIP / middleware.OptionalHeader
directly. Drops the unused "net" import from each.
- slowredirect/fingerprint_handler.go: same treatment for the
AttachFingerprint call site.
No behavior change — middleware.RealIP is byte-identical to what was
previously inlined per generator; verified by re-running the per-package
trigger tests post-refactor.
Swaps the Phase-9 directEventRecorder/mysqlEventRecorder shims for
real event.Service.Record via an eventRecorderAdapter. Constructs
notify.Service with telegram + webhook senders. Spawns the retention
loop goroutine off the main wait group; gracefulShutdown now also
notifySvc.Wait()s for in-flight notifications to drain.
Closes deferred Phase-9 items 2 (real event recorder), 3 (nested
create-tier rate limits), 5 (FingerprintRecorder wiring), 7 (drop
required tag on TurnstileResp).
- cmd/canary/main.go:
* buildEventStack constructs telegram + webhook senders and registers
them on notify.Service. event.Service wired with eventRepo (Store +
StatusWriter), tokenRepo (TokenIncrementer), redis client, notifier.
* eventRecorderAdapter and a single mysqlEventRecorder both delegate
to event.Service.Record(ctx, t.NotifyInfo(), evt) — one path,
same dedup + notify semantics for HTTP triggers and mysql triggers.
* fingerprintRecorderAdapter wraps event.Repository.AttachFingerprint
with the configured window (defaults to 5 min) so the slow-redirect
fingerprint POST writes through.
* spawnRetentionLoop runs event.Service.RunRetentionLoop on the
shared wait group; configured interval + per-token limit gate it.
* /api/tokens POST gains nested create-tier limits per spec §11.3:
LimitCreateMin (5/min) + LimitCreateHour (20/hr) chained ahead of
Turnstile, with their own fingerprint-keyed Redis namespaces so
they don't share buckets with the read-tier limiter.
* Shutdown order: server → telemetry → redis → db, then
notifySvc.Wait() to drain in-flight goroutines, then wg.Wait()
for retention loop + mysql listener.
- internal/config/config.go:
* NotifyConfig: dedup_ttl, send_timeout, max_tries, max_elapsed,
initial_interval, retention_interval, retention_limit,
webhook_hmac_secret, telegram_api_base, fingerprint_window.
* RateLimitConfig: create_min_rate / burst, create_hour_rate / burst.
* Defaults + env mappings (NOTIFY_*, RATE_LIMIT_CREATE_*, plus the
spec'd WEBHOOK_HMAC_SECRET).
- internal/token/dto.go: drop the `validate:"required"` tag on
cf_turnstile_response. The middleware (TurnstileVerify) is the sole
authority — it bypasses on empty SecretKey for dev mode and would
reject empty-token submissions in prod via the Verifier's own
ErrEmptyToken path. The validator tag was forcing tests to pass
a placeholder string in dev mode which conflicted with the bypass.
Audit found that registry.Build was calling mysql.New() with hardcoded
localhost:3306 defaults, ignoring cfg.MySQL.PublicHost and PublicPort.
Result: every emitted mysql:// connection string said
"mysql://canary_<id>@localhost:3306/internal_db" regardless of actual
deployment address — useless for any non-localhost deployment.
Fix:
- registry.Config: adds MySQLPublicHost + MySQLPublicPort fields
- registry.Build: switches between mysql.New() and
mysql.NewWithAddress(host, port) based on config
- cmd/canary/main.go: passes cfg.MySQL.PublicHost/PublicPort
through to registry.Build
- registry_test.go: new TestBuild_MySQLUsesConfiguredPublicAddress
asserts the connection string reflects configured host/port
Other Phase 9 audit observations DEFERRED to Phase 10 (not blocking
rollup):
- §11.3 create-tier rate limit on POST /api/tokens (Turnstile
already provides spam protection; tiered limits are
belt-and-suspenders)
- scoping the gosec G107/G704 exclude to specific packages
(revisit when Phase 10's webhook sender lands and needs
case-by-case URL validation)
- TurnstileResp dto required-tag clashes with dev-mode bypass
(low severity — current behavior is sound)
- /api/m/{manage_id} GET/DELETE (spec §8.1 lines 740-741) — Phase 11
- FingerprintRecorder is nil at startup (slowredirect fingerprint
handler wiring) — Phase 10
- mysql_username metadata write in token.Service (mysql trigger
lookup goes by token ID, not username — low impact)
Pre-rollup gate clean at HEAD:
- go build ./... + go vet ./... clean
- go test -race -timeout=60s ./... all pass (~50 new test cases)
- go test -tags=integration -race -timeout=300s ./internal/token/...
./internal/event/... all pass
- golangci-lint run ./... → 0 issues
- grep -rn "//nolint" --include="*.go" returns nothing
Phase 9 tasks 9.7 + 9.8 + closes Phase 8 task 8.5 deferral.
config.Config additions:
- Canary {BaseURL, ManageURL} — public URLs for trigger
+ manage redirects
- Turnstile {SecretKey, SiteKey} — Cloudflare Turnstile
(empty SecretKey = dev bypass)
- MySQL {Enabled, Addr, PublicHost, — fake TCP listener config
PublicPort}
cmd/canary/main.go full wire-up per spec §6.3 + §11.1:
- Global middleware: RequestID → Logger → Recovery → SecurityHeaders
- /healthz: no further middleware (router-mounted directly)
- /c/{id}, /c/{id}/fingerprint, /k/{id}, /k/{id}/*: mounted at
router root, OUTSIDE /api subrouter, so they skip CORS + rate
limit (spec §11.1 line 1552 — we want to record EVERY hit)
- /api subrouter: CORS → RateLimit (KeyByFingerprint, FailOpen=true)
- /api/tokens/types: no further middleware (read-only public list)
- /api/tokens POST: + TurnstileVerify middleware
- mysql goroutine: spawned via spawnMySQLListener iff
cfg.MySQL.Enabled; uses mysqlTokenLookup + mysqlEventRecorder
adapters bridging mysql.{TokenLookup,EventRecorder} to
token.Service + event.Repository
Adapter types in cmd/canary/main.go:
- registryAdapter: bridges registry.Registry (map type) to
token.Registry interface (Get method)
- directEventRecorder: bridges token.EventRecorder to
event.Repository.Insert + token.Repository.IncrementTriggerCount
(synchronous; Phase 10 swaps in event.Service for dedup + async
notify)
- mysqlTokenLookup, mysqlEventRecorder: same shape for the TCP
mysql handler
webbug refactor (task 9.8):
- Delete local realIP/lastNonEmptyXFF/optionalHeader copies (~30 LOC)
- Use middleware.RealIP and middleware.OptionalHeader instead
- Tests still pass (the IP-precedence suite in webbug_test.go
exercises middleware.RealIP behavior end-to-end via the Trigger
contract)
Pre-rollup gate clean at HEAD:
- go build ./... + go vet ./... clean
- go test -race -timeout=60s ./... all pass
- go test -tags=integration -race -timeout=300s ./internal/token/...
./internal/event/... all pass
- golangci-lint run ./... → 0 issues
- grep -rn "//nolint" --include="*.go" returns nothing
DEFERRED (not in Phase 9 scope):
- Phase 9 task 9.9 integration test (full POST /api/tokens against
testcontainers) — defer to Phase 10's broader integration suite
along with event.Service + notify.Service end-to-end coverage
- Phase 10 brings: event.Service with dedup + async notify, swaps
in for the directEventRecorder/mysqlEventRecorder adapters
- Phase 13 brings: geoip.Service wired into event recording
Phase 9 task 9.5.
Resolved a long-standing import-cycle constraint: the Generator
interface + Artifact + ArtifactKind + TriggerResponse types now live
at internal/token/contract.go (the token package, which both the
service and concrete generators can import). The
internal/token/generators/generator.go file now re-exports these
as Go type aliases for backwards compatibility — concrete generators
(webbug/slowredirect/docx/pdf/kubeconfig/envfile/mysql) keep their
existing `generators.Artifact` / `generators.KindFile` references
unchanged.
This unlocks token.Service.Create, which would otherwise be unable
to type-reference Artifact (since generators imports token for
token.Type/token.Token).
token.Service:
- Create(ctx, req, fp, ip) (*Token, Artifact, error):
- go-playground/validator on the request DTO
- type-specific metadata validation:
- slowredirect → metadata.destination_url required, must be
http:// or https://
- envfile → metadata.include_keys (if present) must be a
subset of {aws,stripe,github,db}; malformed JSON rejected
- others → no extra validation
- registry.Get(type) → ErrUnknownGeneratorType on miss
- generateTokenID: 12-char [a-z0-9] via crypto/rand
- uuid.NewString for manage_id
- call Generator.Generate (which may mutate t.Metadata for
mysql's mysql_username persist)
- repo.Insert; rollback semantics: if generator fails, no
persistence happens
- GetByID / GetByManageID: proxy to repo, return (nil, nil) on
not-found (so handlers can distinguish 404 from real errors)
- IncrementTriggerCount: proxy to repo
- TriggerURL / ManageURL: URL builders so handlers don't duplicate
path joining
- Generator(t Type): registry lookup for handler dispatch
Tests (~15 cases): ID + manage_id format regexes, repo persistence,
generator call, unknown type, validation failure, slowredirect
metadata variants (missing/invalid-scheme/valid), envfile include_keys
(invalid/valid), generator error path skips persistence, distinct IDs
across 30 calls, GetByID nil-nil semantics, URL builders.
Phase 8 third commit. Closes the generator set — all 7 token types are
now registered.
mysql/generator.go:
- Generator with publicHost / publicPort / database fields. New()
returns localhost:3306/internal_db defaults; NewWithAddress(host,
port) overrides for Phase 9 wire-up
- Generate writes mysql_username into t.Metadata (json.RawMessage
merge via setMySQLUsername helper — same defensive map[string]
json.RawMessage pattern as envfile's extractIncludeKeys: malformed
existing metadata is replaced rather than erroring, existing
fields preserved, prior mysql_username overwritten)
- Returns Artifact{Kind: KindConnectionString, ConnectionString:
"mysql://canary_<id>@<host>:<port>/internal_db"} per spec §9.7
lines 1345-1352. baseURL ignored (mysql connection strings don't
use it)
- Trigger(ctx, t, r) returns (nil, nil, ErrHTTPTriggerNotSupported)
sentinel error — mysql triggers fire over the TCP listener (see
server.go/handler.go), not via the HTTP router. Phase 9's router
will not mount any HTTP route to mysql.Trigger; this is defense
for accidental programmatic calls
- Note on spec §9.7 line 1343: `t.Metadata["mysql_username"] =
username` doesn't compile against the real Token.Metadata
(json.RawMessage). The setMySQLUsername helper is the sanctioned
*string-shape adaptation, matching the slowredirect/envfile
metadata patterns
mysql/generator_test.go (~11 cases):
- Type / KindConnectionString / default connection string format /
custom address / canary_ prefix invariant / metadata persistence
on empty token / preservation of existing metadata fields /
malformed-metadata fallback / stale-mysql_username overwrite /
baseURL ignored / Trigger returns ErrHTTPTriggerNotSupported with
nil event + nil response
registry update: cardinality 6 → 7; pending list now empty.
TestBuild_AllSevenGeneratorsRegistered replaces the
TypesPresentInPhaseN test pattern from prior phases, asserting all 7
generator types are present (closes the generator set; future phases
add services/handlers, not generators).
DEFERRED TO PHASE 9 (task 8.5): wiring `go mysql.Run(ctx, addr,
handler)` into cmd/canary/main.go. The handoff anti-relitigation set
documents "Phase 9 wires the registry into main.go" — the mysql
listener goroutine fits naturally into Phase 9's main.go work
alongside config.Config.MySQL.Enabled, config.Config.MySQL.Addr,
the token.Service (for TokenLookup), and the event.Service (for
EventRecorder). Wiring it in Phase 8 would require partial config
additions that Phase 9 then re-touches.
Pre-rollup acceptance gate clean at HEAD:
- go build ./... + go vet ./... clean
- go test -race -timeout=60s ./... all packages pass (~115 tests
added across Phase 8 alone)
- go test -tags=integration -race -timeout=300s pass
- golangci-lint run ./... → 0 issues.
- zero //nolint pragmas anywhere
Phase 8 second commit. TCP listener orchestration + per-connection
business logic.
server.go (~70 LOC):
- mysql.Run(ctx, addr, ConnectionHandler) error
Listens via net.ListenConfig (context-aware), accepts in a loop,
dispatches each connection to handler.HandleConnection in its own
goroutine, waits on context cancellation to close the listener
and drain in-flight connections via WaitGroup
- ConnectionHandler interface (single HandleConnection method) so
tests can substitute a stub without standing up real handler deps
- Logs listener bind + accept errors via slog; suppresses net.ErrClosed
on graceful shutdown so the shutdown path is silent
handler.go (~170 LOC):
- TokenLookup interface (GetByID) and EventRecorder interface (Record)
decouple the handler from Phase 9's token.Service and Phase 10's
event.Service. Production wire-up in Phase 9 (deferred — see Phase
8 rollup notes on task 8.5 deferral)
- HandleConnection writes HandshakeV10 → reads HandshakeResponse41 →
extracts username → strips canary_ prefix → looks up token →
records event with mysql_username + mysql_client_capabilities +
mysql_client_charset in event.Extra (mirrors kubeconfig's
kubectl_* forensic capture) → writes ERR_Packet 1045 → defers
close. 10-second connection deadline.
- Defense-in-depth: usernames without canary_ prefix dropped
silently, unknown tokens dropped silently, lookup errors dropped
silently — no behavior signal to a probing attacker
- Nil EventRecorder tolerated (writes ERR but skips event) — Phase 9
can wire the handler before event.Service exists if needed
Tests (~14 cases):
- server: nil-handler rejected, basic dispatch, 10 concurrent
connections, context cancellation closes listener cleanly,
invalid address rejected
- handler: handshake sent first, non-canary username drops silently,
known token records event AND sends ERR, event.Extra carries
all three kubectl-equivalent fields with correct hex formatting,
unknown token records nothing, lookup error silent, nil
EventRecorder tolerated (still sends ERR), bad auth packet
silent
- Uses real net.Listener + net.Conn pairs (not mocks) for
behavioral fidelity — TCP wire-protocol tests should hit real
sockets
go test -race -timeout=60s passes. golangci-lint clean.
Phase 8 first commit. Pure encoding/decoding layer for the MySQL
wire protocol — no I/O orchestration, no DB lookups, no business
logic. Three primary functions plus utility helpers, all driven by
the byte-exact spec §9.7 lines 1363-1403:
- BuildHandshakeV10(connID, authData[20]) []byte
Server's initial greeting packet (0x0a protocol version,
"5.7.40-canary" server version, random auth-plugin-data split
into 8+12 byte parts, capability flags 0xf7ff / 0x81ff,
utf8mb4_unicode_ci charset, mysql_native_password plugin name).
Wraps in 3-byte LE length + 1-byte sequence ID (0x00).
- ReadClientAuth(r io.Reader) (*ClientAuth, error)
Parses HandshakeResponse41. Returns the username (null-terminated,
extracted starting at byte offset 32 after the 4+4+1+23-byte
fixed header) plus the client's capabilities, max packet size,
and charset for forensic richness. Defensive bounds checking
against ErrInvalidPayload / ErrUsernameMissing / ErrShortPacket /
ErrPacketTooLarge sentinels. 64KB max packet size cap.
- BuildAccessDeniedErr(username, sourceHost string) []byte
Standard MySQL ERR_Packet 1045 with SQL state "28000" and the
exact "Access denied for user '%s'@'%s' (using password: YES)"
message kubectl and the mysql client both display verbatim.
Sequence ID 0x02 (after handshake=0 and client auth=1).
Plus crypto/rand-backed helpers NewRandomAuthData (20 bytes for the
challenge) and NewRandomConnectionID (uint32). Consistent with the
crypto/rand discipline from Phase 0 supplement, Phase 5 (pdf), Phase 7
(envfile).
Tests (~24 cases): byte-exact assertions on packet headers, payload
layout, sequence IDs; round-trip Build/Read for 5 username variants;
error paths (short packet, missing username terminator, empty reader,
oversize packet rejected mid-stream); randomness distinctness (50
calls → near-50 unique values).
Handler + server land in next commit; generator + registry in the
one after. Trigger interface conformance for the mysql Generator will
be a sentinel-error no-op (mysql triggers over TCP, not HTTP — Phase
9 router doesn't mount any HTTP path to mysql.Trigger).
go test -race -timeout=60s ./internal/token/generators/mysql/...
passes, golangci-lint clean.
Fifth entry in the registry map. Cardinality bumps 4 → 5; pending list
shrinks to {envfile, mysql} (2 remaining).
- registry.go: adds the import + token.TypeKubeconfig: kubeconfig.New()
- registry_test.go:
- TestBuild_RegistersKubeconfig added (mirrors RegistersPDF)
- TypeKubeconfig removed from TestBuild_PendingTypesNotYetRegistered
- TestBuild_OnlyExpectedTypesPresentInPhase5 renamed → InPhase6,
cardinality assertion bumped 4 → 5
Pre-audit gate clean at HEAD: go build ./... + go vet ./... + go test
-race -timeout=60s ./... + go test -tags=integration -race -timeout=300s
./internal/token/... ./internal/event/... + golangci-lint run ./...
all pass with 0 issues.
Trigger implementation completes the kubeconfig Generator's interface
conformance. Behavior matches spec §9.5 + §8.5 defense-in-depth:
- ALWAYS returns 403 + Kubernetes-shaped Status JSON (whether the
token is valid or not). Attackers cannot distinguish "valid
token, no permission" from "no such token" by status code or
response shape — both are 403 with byte-identical structure modulo
the resource/verb message slot.
- Status JSON conforms to k8s.io Status object: kind=Status,
apiVersion=v1, metadata={}, status=Failure, reason=Forbidden,
code=403, message=verisimilar real-kubectl error string.
- Message format mirrors kubectl's actual output:
<resource> is forbidden: User "system:anonymous" cannot <verb>
resource "<resource>" in API group "" in the namespace "default"
where <resource> is the last segment of r.URL.Path and <verb> is
derived from the HTTP method (GET/HEAD → list, POST → create, PUT
→ update, PATCH → patch, DELETE → delete, others → list). This is
richer than spec §9.5's hardcoded "list" — kubectl ships the verb
based on the action it's attempting, so a /apis/.../pods DELETE
response saying "cannot list" would tip the attacker off. Method
mapping is forensic gold for the operator (which kubectl action
actually fired) and verisimilar to the attacker.
- Event.Extra captures the kubectl_path / kubectl_method /
kubectl_query / kubectl_ua fields per spec §9.5 line 1234. UA
capture is the forensic prize — kubectl ships its version + arch
in the UA string ("kubectl/v1.30.0 (linux/amd64) kubernetes/...").
- Source-IP triplet (realIP / lastNonEmptyXFF / optionalHeader) is
a content-copy of webbug/docx/pdf — sanctioned duplication per
Phase 2/3/4 anti-relitigation set. Phase 9 middleware extraction
collapses these into one shared helper.
- Nil-token path returns the same 403+Status with verisimilar
message + headers, and nil event so the handler cannot persist
a row with empty TokenID (FK violation) — spec §8.5.
Tests cover (~36 cases including subtests):
- 403 + JSON content-type + cache headers
- Status JSON shape (Kind/APIVersion/metadata={}/Status/Reason/Code)
- Message resource extraction: pods, secrets, single-resource get,
API version probe, trailing slash, healthz probe (6 paths)
- Verb derivation from HTTP method (7 methods, 5 distinct verbs)
- User "system:anonymous" impersonation present
- kubectl_path/method/query/ua in Extra JSON
- Empty query string handled
- Source-IP precedence (11 subcases mirroring docx/pdf for parity)
- Missing UA/Referer → nil pointers
- Nil-token defense-in-depth: still returns 403 + parseable Status,
nil event, response-shape parity with valid-token case (HTTP code
+ JSON kind/version/status/reason/code all identical so an
attacker cannot probe token validity by diffing responses)
Generator now satisfies generators.Generator interface. Registry wire-up
in next commit.
Phase 6 first commit. Generator produces a real-looking kubeconfig
artifact that points kubectl at the canary's /k/{id} endpoint:
- text/template embedded YAML following the spec §9.5 layout (one
cluster + one context + one user, all referencing the same
prod-cluster + svc-backup-reader names for verisimilitude)
- APIServerURL = baseURL + "/k/" + t.ID (trailing-slash safe via
strings.TrimRight)
- Token = t.ID embedded as the user's bearer token — when kubectl
fires a request, the bearer in the Authorization header is exactly
this id, giving us a second lookup path beyond the URL path
- ClusterName = "prod-cluster", UserName = "svc-backup-reader" —
hardcoded per spec §9.5 lines 1193-1194 (tempting bait names that
look like a real prod service-account kubeconfig)
- Artifact.Kind = KindText (different from docx/pdf KindFile — YAML
is text)
- ContentType "application/yaml"
- Filename default "kubeconfig" with the standard *string deref+trim
pattern from docx/pdf
The template file is named template.yaml.tmpl (not .yaml) so the
repo's pre-commit check-yaml hook does not parse it as pure YAML —
the {{.X}} Go-template tokens are not valid YAML flow mappings and
would fail the parser. .tmpl is the conventional extension for
Go-templated source files; identify-cli (which check-yaml uses to
match files) does not classify .yaml.tmpl as YAML. Future phases
needing templated YAML (envfile recipes, etc.) should follow the
same convention.
Tests parse the rendered output through gopkg.in/yaml.v3 against a
typed schema struct, asserting every field flows through correctly:
APIVersion, Kind, current-context, Clusters[0].name + .cluster.server,
Contexts[0].name + .context.cluster + .context.user, Users[0].name +
.user.token. Plus Filename defaulting subtests, base-URL
trailing-slash trim, subpath preserve, distinct-ids → distinct
outputs, bearer-token-embedded check, ends-with-newline check.
Handler (Trigger) ships in the next commit per spec §9.5 file split
({generator, handler, template}). At this commit kubeconfig.Generator
has Type() + Generate() but no Trigger() — the package compiles in
isolation but does not yet satisfy generators.Generator interface.
The registry wires it in commit 3.
go test -race -timeout=60s ./internal/token/generators/kubeconfig/...
passes.
Phase 5 first commit. Adds the committed template.pdf binary blob plus
the pure-Go one-shot utility that produced it, matching the docx Phase
4 pattern (cmd/build<format>template/main.go excluded from production).
cmd/buildpdftemplate/main.go assembles a minimal PDF-1.4 by hand:
- Header: %PDF-1.4 + 4 high-bit bytes to mark as binary
- Object 1: Catalog → Pages root
- Object 2: Pages → single Kids ref + Count 1
- Object 3: Page with /AA << /O << /Type /Action /S /URI /URI (...) >> >>
where the URI value is the 76-char placeholder
HONEY_TRACK_URL_PADDED_TO_FIXED_WIDTH______________________________________
- xref table with computed byte offsets
- trailer + startxref + %%EOF
The placeholder is a direct dictionary value (NOT inside a stream), so
substitution can be a plain byte-replace at runtime without breaking the
cross-reference table — spec §9.4 line 1133. No FlateDecode anywhere.
pdfcpu (v0.12.1) is used for validation only — the builder calls
api.Validate before writing, and tests will call it again on substituted
output to confirm the PDF stays well-formed.
Verification at HEAD:
- 477-byte template.pdf
- sha256: 70c6359016ceb780539f8c4497991b98ff82201e35a38d36090d88856027e103
- Reproducible: rebuild produces byte-identical output
- grep -aob HONEY_TRACK_URL → one offset (240)
- file: PDF document, version 1.4, 1 page(s)
- api.Validate passes
go.mod adds pdfcpu as a direct dep with its transitive deps via go mod
tidy. golang.org/x/crypto returns to direct (pdfcpu uses it for AES
encryption support).
Phase 4 audit (2 agents) returned PASS with three NITs. N1 + N2 cleared
in-phase per the standing fix-in-phase rule. N3 was stylistic-consistency
positive and not actionable.
N1 — generator_test.go IP-precedence subtests: added two missing cases
present in webbug's identical suite (the realIP helper triplet is byte-
identical to webbug, so this is symmetry rather than coverage):
- "RemoteAddr loopback IPv6 strips brackets and port" → "[::1]:9999"
- "XFF IPv6 rightmost" → "198.51.100.1, 2001:db8::dead"
N2 — generator.go patchTemplate loop: the second `cErr := ...` shadowed
nothing but reused the name from the close-error block earlier in the
same iteration. Renamed the header-create variant to `hErr` for clarity.
Audit verdicts:
- superpowers:code-reviewer: PASS (0 B / 0 S / 3 N)
- general-purpose spec-adherence vs §9.3 / §8.5: PASS (0 violations,
6/6 invariants MATCH, BACKLOG.md still empty)
Phase 4 Task 4.3+4.4 — docx Generator that produces a tracked Word
document by embedding template.docx and runtime-substituting the
HONEY_TRACK_URL placeholder in word/footer2.xml with the per-token
trigger URL. Preserves per-entry zip Method (STORE/DEFLATE) so
Word/LibreOffice still opens the result; non-footer entry bodies are
byte-identical to the template.
Generate returns {Kind: KindFile, Filename: t.Filename ?? "Document.docx",
Content: <patched zip>, ContentType: wordprocessingml MIME}. The
spec §9.3 reference snippet uses `t.Filename == ""` which won't
compile against the real schema — Token.Filename is *string. The
generator dereferences safely via resolveFilename, trimming and
falling back to the default for nil/empty/whitespace pointers.
Trigger mirrors webbug exactly per spec §9.3 line 1118: 200 + 43-byte
transparent GIF via pixel.Clone() + cache-control no-store + pragma
no-cache. Nil-token returns the same response with nil event
(spec §8.5 — no token-existence enumeration). realIP /
lastNonEmptyXFF / optionalHeader helpers are copied from webbug per
the standing rule (sanctioned duplication until Phase 9 middleware
extraction).
24 test cases including the six prescribed by implementation plan
§4.3 (OutputIsValidZip, FooterContainsTriggerURL,
FooterDoesNotContainPlaceholder, OtherEntriesUnchanged,
PreservesCompressionMethods, ReturnsGIFLikeWebbug) plus surrounding
correctness/regression coverage (type, kind, content type, filename
defaulting incl. whitespace, trigger-URL trailing-slash + subpath +
uniqueness, source-IP precedence with 9 subtests mirroring webbug,
GIF body independence per Trigger call, nil-token defense).
Also fixes pre-existing lint debt in cmd/builddocxtemplate from
commit f36cce9a: errcheck on the deferred f.Close (now propagates
via named return) and gosec G304 on os.OpenFile (now filepath.Clean
on the operator-supplied -out path). Rebuilding template.docx after
this change produces byte-identical output (verified via sha256sum).
Phase 4 Task 4.1+4.2 — pure-Go OOXML template builder under
backend/cmd/builddocxtemplate/. Run:
go run ./cmd/builddocxtemplate \
-out ./internal/token/generators/docx/template/template.docx
produces a minimal valid .docx (5 zip entries, mixed STORE/DEFLATE
compression) whose word/footer2.xml carries an INCLUDEPICTURE field
referencing the literal placeholder string HONEY_TRACK_URL with the
\\d switch (forces fetch-on-open, no local cache).
The committed template.docx is the embedded source for the Phase 4
docx generator (next commit). Mixed compression methods in the template
make Task 4.3's TestGenerate_PreservesCompressionMethods a meaningful
regression guard — a hardcode-DEFLATE generator would now mismatch on
[Content_Types].xml + word/_rels/document.xml.rels (both STORE).
cmd/builddocxtemplate is excluded from the production binary (separate
cmd/ subdir from cmd/canary).
Phase 3 audits returned one BLOCKER and one SPEC-VIOLATION plus two
should-fix items; all cleared in-phase per the no-rot rule.
BLOCKER — template.html double-escaped JS-context interpolations.
Go's html/template auto-escapes {{.X}} in <script> string-literal
position via its built-in jsstrescaper. Piping through the explicit
`| js` (JSEscaper) re-ran the escaper, producing `\\u003D` /
`\\u0026` on the wire (double backslash). JS parses `\\u003D` as a
literal backslash followed by "u003D" text rather than `=`, so any
realistic destination URL — e.g.
https://news.example.com/article?utm_source=newsletter&utm_medium=email
— was corrupted into
https://news.example.com/article?utm_source=newsletter&utm_medium=email
before being passed to window.location.replace. The unit tests
previously missed it because every test destination was an
escape-free string. Fix: drop `| js` from both JS-context actions
in template.html; the contextual auto-escaper alone produces single-
backslash `=` which JS correctly decodes. Spec §9.2 line 1031
and 1038 (and the explanatory paragraph 1045) are factually wrong
about `| js` being a custom template func — it is `JSEscaper`, and
adding it on top of the built-in contextual JS string escaper is
pure double-escape harm. Spec docs (local-only, not committed)
amended accordingly. Added
TestTrigger_DestinationRoundtripsThroughJSStringDecode as a
regression guard against re-introducing the pipeline.
SPEC-VIOLATION — nil-token Trigger returned `404 + "Not Found"`,
breaking spec §8.5 ("token not found | 200, ... deliberately do NOT
404") and §8.5 line 964 ("/c/* endpoints never return JSON errors").
A 404 on a non-existent token ID lets scanners enumerate which IDs
are real. Fix: nil-token branch now renders the same template with a
benign decoy destination ("/") and returns 200 + text/html with the
full CSP override and cache headers. Response is byte-shape
indistinguishable from a valid-token response except for the
destination URL embedded in the script. Added
TestTrigger_TokenNotFound_HasSameResponseShapeAsValidToken to
assert the indistinguishability invariant. The `(nil event, &resp,
nil error)` return shape from the handoff anti-relitigation is
preserved.
SHOULD-FIX S1 — extractDestination now allowlists `http://` and
`https://` (case-insensitive) and returns the new
ErrInvalidDestinationScheme sentinel otherwise. The noscript
meta-refresh sits in an HTML-attribute context where html/template
does NOT recognize `url=...` as a URL context, so `javascript:`,
`data:`, `file:`, `vbscript:` and scheme-relative `//example.com`
would otherwise render verbatim and follow on JS-disabled clients.
Phase 9's `validate:"url"` on input is necessary but not sufficient
(the validator's `url` tag accepts any scheme). Added
TestGenerate_RejectsDangerousDestinationSchemes (8 cases) and
TestGenerate_AcceptsHTTPAndHTTPSSchemes (4 case-variants).
SHOULD-FIX S3 — fingerprint handler now silently 204s on any
Content-Type that does not start with "application/json". The
embedded template is the only legitimate caller and always sends
JSON; this rejects adversarial probes before the MaxBytesReader
read. Added
TestFingerprintHandler_WrongContentType_Returns204AndDoesNotTouchEvent.
SHOULD-FIX S4 — added integration tests for the oversize-body cap
(128 KiB body returns 204, Extra untouched) and the empty-token-id
guard (`/c//fingerprint`); the latter accepts chi's natural routing
behavior alongside the in-handler 204.
Non-functional cleanups: extracted shared template-render path
into renderWith / renderResponse / renderDecoyResponse to keep the
nil-token + happy-path bodies definitionally identical;
ErrMissingDestination and ErrInvalidDestinationScheme exported so
callers (and tests) can ErrorIs them; TestGenerate_RejectsMissing*
tightened to use require.ErrorIs(tc.wantErr) instead of bare
require.Error.
Pre-rollup gate (go build / vet / test -race unit+integration /
golangci-lint) is clean.
Phase 3 wiring: registry.Build now maps token.TypeSlowRedirect to
slowredirect.New() alongside the existing webbug entry. The pending-
types regression guard sheds SlowRedirect from its list (5 remain:
docx, pdf, kubeconfig, envfile, mysql) and the cardinality assertion
moves from "exactly one generator" to "exactly two".
POST /c/{id}/fingerprint handler that decodes a JSON fingerprint body
and merges it into the most recent event's Extra JSONB for the same
(token_id, source_ip) tuple within a 30s window via the existing
event.Repository.AttachFingerprint method. Decoupled from the concrete
repository through a small FingerprintAttacher interface so future
event.Service can swap in without touching this package.
Per spec §9.2 the fingerprint POST is enrichment-only: a missing match
(event.ErrNotFound), an invalid JSON body, an empty body, and a body
exceeding 64 KiB all silently return 204. Only unexpected repository
errors are logged via slog. The token id is read from chi.URLParam("id")
so the route mounts naturally under the existing trigger router in
Phase 9.
Three integration tests against testcontainers cover the happy-path
JSONB merge, the no-matching-event path (silent 204), and the
invalid-JSON path (204 plus stored Extra left untouched).
HTML+JS browser-fingerprint redirect page rendered via html/template:
Generate reads metadata.destination_url off the token row and returns
KindURL with the trigger URL + persisted destination; Trigger renders
the embedded template (noscript meta-refresh + inline fetch posting
the fingerprint, then window.location.replace), returns the page with
Content-Security-Policy override allowing inline script + connect-src
'self', and emits the event capturing token id / source IP / UA /
Referer.
Nil-token Trigger returns nil event (FK guard) plus a 404 HTML body,
matching the established defensive-rendering pattern from webbug.
Destination_url is extracted defensively (missing/empty/whitespace
all surface ErrMissingDestination). realIP precedence copied from
webbug pending the Phase 9 middleware promotion.
19 unit-test assertions cover Type, Generate's URL+destination output,
five missing-destination edge cases, XSS neutralization in two
injection contexts (attribute escape + script-closing escape), CSP
override + cache headers, event-recording metadata, nil-token contract,
and per-call response independence.
Two audit agents (code-reviewer + spec-adherence) both returned PASS but
flagged substantive findings. Per fix-in-phase / no-backlog-rot, clearing
every MEDIUM + LOW in-phase.
Findings addressed:
MEDIUM — realIP RemoteAddr fallback returned "IP:port" verbatim, where
the spec wants the bare IP (geoip + downstream parsing assume host-only).
Now uses net.SplitHostPort with raw-string fallback if not host:port.
Adds 4 RemoteAddr cases: IPv4 strips port, IPv6 bracket form strips
brackets+port, loopback IPv6, and the port-less raw fallback.
MEDIUM — realIP XFF branch accepted whatever rightmost-comma-split
produced, so headers like "198.51.100.1, " (trailing comma) yielded
"" and skipped XRI/RemoteAddr entirely. Extracted into lastNonEmptyXFF
which walks right-to-left and falls through cleanly. Adds two cases:
trailing-comma falls through, all-empty entries fall through.
MEDIUM — pixel.TransparentGIF was an exported mutable []byte; any caller
could clobber it process-wide (mutating one response's body would affect
every subsequent webbug trigger). Renamed to unexported transparentGIF
+ exposed pixel.Clone() and pixel.Len(). Webbug now calls pixel.Clone()
per trigger; new test TestTrigger_ResponseBodyIsIndependentCopyPerCall
verifies two triggers return independent slices.
LOW (reviewer-escalated to HIGH-for-Phase-3) — Trigger on nil token
returned a non-nil event with empty TokenID. Since events.token_id is a
NOT NULL FK to tokens.id, the Phase 3 handler could not have persisted
that event anyway, and the contract was implicit (would have required
an inline comment, which the no-comments rule forbids). Now: nil token
in → nil event, non-nil response out. Contract is explicit in the
return shape. Test asserts evt is nil for nil-token path.
LOW — no IPv6 coverage in IP-precedence tests. Added IPv6 cases for
XFF rightmost and three RemoteAddr forms (bracketed, loopback, no-port).
NIT — gracefulShutdown swallowed every shutdown error, returning nil
unconditionally. Now collects with errors.Join so callers can detect
partial-shutdown for telemetry/exit-code purposes.
Audits accepted as-is (NIT, not blocking):
- Artifact discriminated-union shape (Phase 3+ concern when other Kinds
are produced)
- Registry returning bare map (read-only post-Build; concurrent reads
are safe by Go's memory model)
- Build(_ Config) ignoring its arg (signature reserved for future
stateful generators)
Quality verified post-fix: build/vet/lint clean (0 issues),
14 webbug tests + 6 pixel tests + 4 registry tests PASS under -race,
integration tests unchanged.
Pre-audit gate (`golangci-lint run`) at the start of Phase 2 surfaced 15
issues that should have been zeroed before Phase 1 rollup. Per the
fix-in-phase rule (no backlog rot), clearing everything before Phase 2
audit agents run on a green tree.
Config:
- .golangci.yml: migrate `issues.exclude-rules` and `issues.exclude-dirs`
to v2 syntax (`linters.exclusions.{rules,paths}`); test-file funlen/
dupl/goconst exclusion now actually applies under golangci-lint v2.10
- .golangci.yml: add G706 to gosec excludes — false-positive log-injection
reports on slog structured-logging call sites (slog separates message
from kv args, immune to log-line injection by construction)
errcheck (5):
- cmd/canary/main.go: `_ = telemetry.Shutdown(...)` → log on error
- internal/token/repository.go: `defer stmt.Close()` → log on close error
- internal/event/repository.go: same as token
- internal/testutil/postgres.go: `_ = pgContainer.Terminate(...)` and
`_ = db.Close()` → t.Logf on cleanup error (use distinct err names to
avoid govet shadow)
funlen (1):
- cmd/canary/main.go: split `run` into `run` + `initTelemetry` +
`mountRouter` + `gracefulShutdown` (was 55 statements, now under
the 50 cap; helpers are individually well under)
govet shadow (1):
- cmd/canary/main.go: migrations check switched from `if err :=` to
`if err =` (reuses outer err whose value was already consumed) — the
inner short-decl was lexically shadowing without intent
- cmd/canary/main.go: select-case `err` renamed to `startErr` to avoid
the same shadow pattern
golines (6, all auto-fixed by `golangci-lint run --fix`):
- main.go, config.go, telemetry.go, event/repository.go, token/dto.go,
token/repository.go — long lines wrapped to 80-col
Verified post-fix: build OK, vet OK, unit tests PASS under -race,
integration tests PASS under -tags=integration -race (12s testcontainers),
golangci-lint reports 0 issues.
CarterPerez-dev