0.9.0 - Pipeline Investigation Tools

Pipeline investigation tools - three new MCP surfaces (get_pipeline_summary, get_job_log_smart, list_pipeline_jobs extension) for AI-agent-friendly CI failure investigation. Closes #64 (the reincarnation arc through #86 → #99). Substantial review cycle: 3 contributor rounds + 1 maintainer Path-B reincarnation + 5 codex bot review rounds + 2 pre-push agent passes closed 21 real bugs end-to-end.

Contributors: @ecthelion77 (Olivier Gintrand) authored the original three-tool design + initial implementation + three response rounds; maintainer added contract-clarity / silent-failure / perf hardening on top via Path B reincarnation (#99).

Added

• get_pipeline_summary tool - single-call pipeline investigation returning
  pipeline details, jobs grouped by stage, and log tails for failed jobs. Includes
  failure pattern detection and max_failed_jobs_with_logs cap (default: 5) to
  prevent context blowup. (#64)
• get_job_log_smart tool - intelligent log post-processor that strips ANSI
  codes, GitLab timestamps, and section markers. Supports tail/head, section
  extraction, and error_only filtering. (#64)
• list_pipeline_jobs extension - new optional include_log_tail,
  log_tail_lines, and max_log_tail_jobs parameters. When include_log_tail=true,
  failed jobs include their cleaned log tail directly in the response. (#64)
• truncated field in get_pipeline_summary response - signals when pagination
  cap (50 pages) was hit and job list may be incomplete.
• section_matched and error_lines_matched fields in get_job_log_smart
  response - explicit feedback on filter effectiveness.
• log_fetch_errors field in pipeline summary - surfaces per-job log fetch
  failures instead of silently swallowing them.

Changed

• Renamed log_lines → log_tail_lines in get_pipeline_summary schema for
  consistency with list_pipeline_jobs. (#64 review)
• failure_pattern is now an exhaustive discriminated union
  (no_failures | single | shared_reason | mixed | unknown) instead of a
  plain string. Both shared_reason and mixed variants carry
  unreasoned_count so callers can tell when some failed jobs lacked a
  failure_reason (the invariant sum(reasons) + unreasoned_count === failedJobs.length holds for both). (#64 review + #99 codex)
• Stage status derivation now mirrors GitLab's aggregation: allow_failure
  jobs no longer poison the stage; mixed success+skipped collapses to success.
  (#64 review)
• Pagination no longer relies on X-Total header (absent in GitLab EE);
  uses items.length < per_page with a MAX_PAGES=50 safety cap. (#64 review)

Fixed

• error_only mode in get_job_log_smart now correctly returns an empty log
  (instead of the full log) when no error-like lines are found. (#64 review)
• Silent log fetch failures - getJobLogTails and getPipelineSummary now
  track and report errors per job instead of empty catch blocks. (#64 review)
• E2E tests - assertions moved outside try/catch so test failures propagate
  correctly; try/catch narrowed to the fetch step only. (#64 review)
• list_pipeline_jobs (include_log_tail=true) now always emits the
  unified {jobs, log_fetch_errors?} wrapper, including the zero-failed-jobs
  fallback path that previously reverted to the legacy flat-array shape.
  (#64 round-3 follow-up)
• get_job_log_smart returns line_count: 0 for an empty log instead of
  the JS-quirk 1 from ''.split('\n'). Faithful counts on the
  section-not-found and error_only no-match paths. (#64 round-3 follow-up)
• FailurePattern.single.reason normalizes empty-string failure_reason
  to null (was passed through verbatim by ??). Aligns with .filter(Boolean)
  semantics used by the multi-job variants. (#64 round-3 follow-up)
• FailurePattern.mixed gains unreasoned_count for parity with
  shared_reason. Previously the mixed branch silently dropped failed jobs
  whose failure_reason was missing/empty from the visible reason histogram.
  (#99 codex)
• get_job_log_smart section extraction is now exact-match. The previous
  regex section_start:\d+:NAME[^\n]*\n? accepted a prefix match - requesting
  section: "build" would silently return build_extra content with
  section_matched: true. Now uses a lookahead (?=[\r\n\[]|$) after the
  name to require GitLab's actual section delimiters. (#99 codex P2 round-2)
• stripSections now consumes both CR and LF after the marker. GitLab's
  section line section_*:NNN:name\r\x1B[0K\n collapses to
  section_*:NNN:name\r\n after ANSI stripping; the previous regex tail
  [\r\n]? consumed only one of CR/LF, leaving an orphan \n per marker.
  Cleaned logs had spurious blank lines and tail: N could shift by an
  empty line. New regex tail \r?\n? consumes the full CRLF combo.
  (#99 codex P2 round-3)
• sections_found surfaces the bare section name. Previously emitted
  "script[collapsed=true]" verbatim, which is unusable as a follow-up
  section: argument (end markers never carry the [option] suffix and
  would never match). Now strips the [...] collapsed-marker tail after
  ANSI removal. (#99 codex P2 round-4)
• pipeline_id: 0 no longer silently falls through to "latest pipeline".
  Schema now requires z.number().int().positive() and the runtime path
  selector uses !== undefined. Previously a JS truthy check disagreed
  with the schema's XOR refine. (#99 pre-push silent-failure-hunter)
• section: "" no longer silently skipped. Schema now requires
  z.string().min(1); empty-string requests fail at parse instead of
  silently returning the full log with section_matched: null.
  (#99 pre-push silent-failure-hunter)
• section matching is now case-SENSITIVE. Dropped the regex i flag.
  GitLab section names are identifiers and case-insensitive matching
  would create silent ambiguity between sibling sections like Build
  and build. Aligns with the "exact-match" contract introduced in the
  codex P2 round-2 fix. (#99 pre-push silent-failure-hunter LOW-2)
• Section markers are stripped regardless of strip_ansi. Previously
  section-marker stripping was bound to the strip_ansi: true path; a
  caller debugging raw ANSI output would see raw section_start:NNN:name
  bytes leak into tail/head windows. Section markers are unconditionally
  noise and are now always stripped. (#99 pre-push code-simplifier)

Added

• log_fetch_capped field in get_pipeline_summary summary and in
  the list_pipeline_jobs + include_log_tail wrapper. Present as
  { fetched: number; total_failed: number } when the helper intentionally
  skipped log-fetching for failed jobs beyond the cap
  (max_failed_jobs_with_logs default 5, max_log_tail_jobs default 10).
  Closes a silent fallback where callers saw a partial set of log tails
  without any signal that more failed jobs existed. (#99 pre-push
  silent-failure-hunter)

Fixed (round 5)

• tail: 1 on a single-line log no longer returns "". The previous
  logTail walked all \n from the end including the trailing
  terminator, so tail: 1 on "ERROR\n" returned the empty string and
  log_tail_lines: 50 returned only 49 real lines. logTail/logHead/
  countLines now treat a single trailing \n as the line terminator
  via an endIdx bound, preserving the O(tail) memory promise.
  (#99 codex P2 round-5)
• stripSections now consumes \x1B[0K clear-control bytes between
  CR and LF. When strip_ansi: false, the section-stripping path runs
  on raw GitLab markers section_*:NNN:name\r\x1B[0K\n - the previous
  regex tail \r?\n? left orphan \x1B[0K\n fragments in the cleaned
  log. New tail \r?(?:\x1B\[[0-9;]*[a-zA-Z])*\n? consumes any inline
  ANSI escapes between CR and LF, working whether ANSI was pre-stripped
  or not. (#99 codex P2 round-5)
• get_job_log_smart.log is shape-stable across the truncation
  boundary. The same tool no longer returns "L\n" for tail: 50
  (no truncation) and "L" for tail: 49 (truncation kicks in) on the
  same 50-line input. Unconditional strip of a single trailing \n at
  the end of the helper. (#99 pre-push silent-failure-hunter)

Performance

• Job log tail/head/section/error_only extraction in
  get_pipeline_summary, get_job_log_smart, and list_pipeline_jobs + include_log_tail is now O(K) in memory instead of O(N) - the previous
  log.split('\n').slice(-N).join('\n') pattern allocated an array of all
  lines just to keep the last/first K. New private helpers (logTail,
  logHead, countLines) walk newlines via lastIndexOf / indexOf /
  charCodeAt. Section extraction uses RegExp.lastIndex instead of
  rawLog.slice(startIdx) to avoid copying multi-MB log strings.
  error_only filtering uses a single gim regex match instead of
  split + filter + join. (#99 gemini)

Security

• Schema parameters now have strict .int().min().max() bounds to prevent
  abuse (e.g., log_tail_lines capped at 200, max_log_tail_jobs at 20).
• XOR validation via .refine() on mutually exclusive parameters
  (pipeline_id/ref, tail/head).