Files
Gitea-Tools/docs/architecture/glitchtip-readonly-tools-design.md
T
sysadmin 079eb60e30 docs: GlitchTip read-only error/event tools design (#73)
Add docs/architecture/glitchtip-readonly-tools-design.md: implementation-
ready design for the GlitchTip observability boundary's read-only tool set —
minimum tools (whoami, list_projects, list_unresolved, get_issue,
recent_events, search; GET-only), field-level allowlist projection with an
explicit never-returned list (headers, cookies, auth fields, emails, IPs,
bodies, locals, raw frames), default fingerprint/release/summary+permalink
output, raw access gated behind a distinct approval-only
glitchtip.event.read_raw operation, cursor pagination with explicit
truncation flags, per-service credential/profile requirements (token by
reference, fail closed), Sentry-compat API subset note, fail-closed failure
matrix, mocked testing strategy with negative PII assertions, and a
readiness checklist. Namespace left to ADR-0001 owner decision #2.

Design only; no implementation, no mutation tools, no code behavior changed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-02 15:23:20 -04:00

175 lines
8.2 KiB
Markdown

# GlitchTip Read-Only Error/Event Tools — Design Notes
- **Status:** Design (implementation-ready notes; **no implementation in this repo**)
- **Issue:** #73 (umbrella: #75; boundary decision: ADR-0001, #71)
- **Related:** #74 (GlitchTip→Gitea filing workflow — composes these read tools),
#78 (dedup/linking, child of #74), #76 (per-service profile schema)
- **Date:** 2026-07-02
## 1. Purpose and scope
Define the minimum **read-only** GlitchTip MCP tool set that lets an LLM answer:
*"What unresolved errors does project X have (by environment/release), and what
is this specific error?"* — with privacy-safe output suitable for LLM context,
issue bodies, and audit logs.
Strictly read-only, per ADR-0001:
- **No mutation tools** — no resolving/ignoring/assigning issues, no comment
posting, no project/team/key administration, no deletes.
- **No automatic GlitchTip→Gitea filing** (that is #74's *orchestrated,
explicitly-invoked* workflow; it composes these read tools and Gitea write
tools — never one dual-credential server).
- **This server never holds Gitea write credentials.**
## 2. Boundary placement (namespace pending)
These tools belong to the GlitchTip observability boundary of the MCP Control
Plane — `glitchtip-mcp` (ADR-0001's recommendation), `observability-mcp`, or
folded into `ops-mcp`. **ADR-0001 open owner decision #2 picks the name; this
design does not assume it.** Tool names below use the `glitchtip_` prefix for
readability and rename mechanically with the decision.
Fixed regardless of the name (per `tool-boundaries.md`,
`credential-isolation.md`):
- Own server process, own `.env`, GlitchTip credentials only.
- No Gitea, Jenkins, or Ops tokens in this runtime; no GlitchTip token
anywhere else.
## 3. API surface note (Sentry compatibility)
GlitchTip implements a Sentry-compatible REST API (`/api/0/...` — organizations,
projects, issues, events). The design targets **GlitchTip's documented subset**
only; Sentry-only endpoints must not be assumed. The implementation should pin
against a tested GlitchTip version and treat missing endpoints/fields as
degraded-but-safe (omit field, never crash).
## 4. Minimum read-only tool set
| Tool | Purpose |
|---|---|
| `glitchtip_whoami` | Verify authenticated identity + active profile (mirror of `gitea_whoami`; fail-closed identity proof) |
| `glitchtip_list_projects` | Projects visible to the token (org-scoped), with pagination bounds |
| `glitchtip_list_unresolved` | Unresolved issues for a project, filterable (§6), sorted by last-seen |
| `glitchtip_get_issue` | Safe detail of one issue (fields §5) |
| `glitchtip_recent_events` | Recent events for an issue (summaries only, §5) |
| `glitchtip_search` | Issue search within a project (query + filters §6) |
All tools are `GET`-only. No tool issues PUT/POST/DELETE.
## 5. Privacy: field-level allowlist (the core rule)
Error events routinely contain PII and secrets (request bodies, cookies,
headers, tokens, user emails/IPs, local variables). Therefore: **allowlist
projection only — raw event/issue payloads are never passed through.**
### Issue-level safe fields (`glitchtip_list_unresolved`, `glitchtip_get_issue`, `glitchtip_search`)
| Field | Notes |
|---|---|
| `issue_id` | GlitchTip issue ID (dedup key for #78) |
| `fingerprint` | When available (dedup key for #78) |
| `title` / `culprit` | Error type + short message/transaction — redactor-passed |
| `project` | Slug |
| `level` | error/warning/… |
| `status` | unresolved/… |
| `environment` | When filtered/available |
| `release` | Version string |
| `first_seen` / `last_seen` | ISO-8601 UTC |
| `event_count` / `user_count` | Numbers only — never user identities |
| `permalink` | GlitchTip web URL (the "link, not dump" principle) |
### Event-level safe fields (`glitchtip_recent_events`)
`event_id`, `timestamp`, `level`, `environment`, `release`, redactor-passed
`message`, and a **stack summary** only: top N (default 5) frames as
`module/filename:function:line` — in-app frames preferred.
### Redact / omit — never returned
Request headers; cookies; auth/session fields; user emails, usernames, IPs;
request/form bodies; query strings; local variables; full raw stack frames
(source context lines); SDK/device metadata beyond platform name; breadcrumbs;
any `extra`/`context` blobs.
Full raw frames or request context require a **separate, explicitly approved**
operation (`glitchtip.event.read_raw`) that is absent from default profiles —
same pattern as `jenkins.console.read` in the #72 design. Even then, output
passes the shared secret redactor; redaction failure ⇒ error, never raw text.
**Default output = fingerprint / release / summary + permalink.** The
permalink carries the human to the full data in GlitchTip's own UI, where its
access control applies — the MCP layer does not re-serve raw payloads.
## 6. Filtering and pagination
Filters (all optional, combinable): `project` (required for issue/event
queries), `environment`, `release`, `fingerprint`, free-text `query`
(GlitchTip search syntax, e.g. `is:unresolved`).
Pagination: cursor-based per the API. Bounds: per-page cap 50; default overall
cap 100 items; hard cap `max_pages` (default 10) against runaway loops —
mirroring `gitea_auth.api_get_all`. Truncation is **explicit** in the return
(`"truncated": true`) — never silent.
## 7. Credentials and profile requirements
Per-service profile model (`gitea-execution-profiles.md`, extended by #76):
- Env/config: `GLITCHTIP_URL`, `GLITCHTIP_ORG`, `GLITCHTIP_TOKEN_SOURCE_NAME`
(secret **name** only; value resolved at runtime, never logged/committed).
- Profile: e.g. `glitchtip-readonly` with namespaced
`allowed_operations: ["glitchtip.read", "glitchtip.event.read"]`
(+ `glitchtip.event.read_raw` only with explicit approval);
`forbidden_operations: ["glitchtip.issue.mutate", "glitchtip.admin"]`
belt-and-braces though no mutating tool exists.
- Missing URL/org/token/profile ⇒ **fail closed** before any network call.
- Read-only ⇒ no confirmation gates; identity (`glitchtip_whoami`) must work so
workflows can prove which account they read as.
## 8. Failure behavior (fail closed, clear, safe)
| Condition | Behavior |
|---|---|
| Unknown project/issue | Explicit `{"found": false, ...}` — no fuzzy matching |
| GlitchTip unreachable (DNS/timeout) | `"network error contacting GlitchTip: <redacted reason>"` — mirror `gitea_auth.api_request` conversion |
| 502/503/504 | "GlitchTip upstream unavailable" |
| 401/403 | "GlitchTip auth failed / insufficient permissions" — no credential echo |
| 429 | Honor Retry-After with capped jittered backoff (as `gitea_auth`) |
| Malformed JSON | "malformed JSON response from GlitchTip" — no raw-body dump |
| Missing profile/creds | Fail closed before any network call (§7) |
All error text passes the shared secret redactor.
## 9. Testing strategy (mocked; for the implementing package)
Mocked-GlitchTip unit tests only, per `docs/developer-testing-guidelines.md`:
- Assert method is always `GET`; URL/filter/cursor shape correct.
- **Projection tests:** response fixtures containing emails, IPs, cookies,
headers, request bodies, locals, full frames ⇒ none appear in output
(explicit negative assertions per §5's redact list).
- Stack summary: top-N frame cap enforced; source-context lines absent.
- Pagination: per-page/overall/max-pages caps; explicit `truncated` flag.
- Filters: environment/release/fingerprint/query passed through correctly.
- Failure matrix of §8 incl. no-token-in-error assertions.
- Profile gate: missing/insufficient profile ⇒ no network call
(`mock_api.assert_not_called()` pattern).
- `read_raw` op absent ⇒ raw-frame request refused without an API call.
## 10. Implementation-readiness checklist
Ready to implement once:
1. ADR-0001 owner decision #2 (namespace/placement) is made — mechanical
rename of the `glitchtip_` prefix if needed.
2. ADR-0001 owner decision #1 (repo home) is made.
3. #76 profile schema exists (or a minimal `glitchtip-readonly` profile is
hand-rolled to the same rules).
4. A pinned GlitchTip version is chosen for API-subset testing (§3).
Explicitly **not** unlocked by this document: any GlitchTip mutation, any
automatic Gitea filing (#74 designs that as a gated, explicitly-invoked
orchestrated workflow), any Gitea credentials in this boundary.