Post-Mortem
Explain why the failure escaped or grew, then measure whether each corrective
control reduces recurrence, detection, or impact risk. Take the code-level
cause from debugging; do not work it out again. Close on controls that are
deployed and falsified, never on reflection. For a process-only failure, use
the observed trail of events and decisions; do not invent a code defect to
enter this skill.
When to use
- A failure reached users or production, escaped far downstream, caused real
loss, or exposed a process failure worth correcting.
- Skip an ordinary reproduced bug. For a technical incident, establish the
technical cause and immediate containment first: invoke
debugging if the
cause is unknown, containing-an-incident if impact is ongoing.
Step 1: Control the record
- Open
assets/post-mortem-template.md when starting the analysis. Fill
Control and evidence handling first.
- Fill
Summary and impact and Timeline from artifacts (any containment
record first), not from memory. Mark every time observed or estimated.
Step 2: Find the structural cause
- Fill
Root cause and contributing conditions: cite the technical cause
from debugging when applicable, or the evidence of the process failure,
then each condition that made the introduction or the impact more likely.
Name conditions, never a person. When there are several conditions, keep
them separate.
- Fill
Escape-path audit: freeze the inventory of gates and surfaces live
when the failure escaped, with the hash of its source, then fill every
column for each entry, and for each omission record a decision that
someone is accountable for.
- Fill every bullet of
Risk-reduction claim for each action. Never claim
that recurrence is impossible.
- Fill the
Corrective actions rows, each mapped to a structural cause;
receipts and External state stay pending and live in the ledger.
- Challenge the draft. Fill
assets/escape-path-challenger.md before
dispatching it read-only as dispatching-parallel-agents Step 2 says.
Without a challenger, record the omission the Expected reviewers field
requires, approved if the failure reached users or caused loss; a
self-review never stands in. Record a decision on every finding. A
findings verdict blocks issuing the analysis until the corrected draft
is challenged again, twice at most; findings still open then go to the
approver.
- Issue the analysis, which is never edited after issue, with its
External lifecycle ledger (the ledger: add to it, never rewrite it), to
the user-set path or the template's path, under the storage authority held
now, and record its identity as the template says. If the record is in the
repository, it is a new state under check.
Step 3: Propose and get the actions accepted
Record each corrective action's proposed state in the ledger.
Present:
Post-mortem {{identity}} — {{path}}
Impact: {{one line}} Structural cause: {{one line}} Escaped gate: {{one line}}
Actions: {{each with owner and date}}
1. Accept the actions
2. Request changes
3. Reject the analysis
4. Cancel
Recommendation: {{the answer the evidence and ownership support}} — {{one sentence}}.
Ask this and end the turn; clarifying-intent owns what closes it.
Move an action out of proposed only on a complete trusted receipt (proof,
from a trusted source outside the state under check, that an action or
decision really happened) for the exact scope and dates. Option 2 issues
a successor analysis and asks again; 3 and 4 record each action rejected
or cancelled, with its residual risk.
Step 4: Prove and track the controls
- Falsify every corrective gate you implement. Record the targeted
fail-then-pass proof in the ledger as raw evidence, never in the issued
analysis: it fails on the captured incident or a representative bad
case, it passes on the good control, and the version before the fix did
not catch it.
- Track each control to a receipt bound to an identity, from the surface
that enforces it: CI, runtime, review, release, alerting, recovery. Never
infer from a summary.
- REQUIRED SUB-SKILL: invoke
verification-strategy for any change to
the battery itself (the set of checks the project runs). Add no duplicate
check nobody owns.
- At the date declared in advance, compare the baseline against the target
and confirm that the gate ran. Write
effective, ineffective, or
inconclusive. If it is ineffective or inconclusive, reopen, unless
the exact approver rule accepts closure with the residual risk.
Action states
Record in the ledger, outside the issued analysis, each action's
move from proposed to one of {owner-accepted | rejected | cancelled | superseded}, and, for an accepted action, from there through implemented,
falsified, deployed/enforced, effectiveness reviewed, then effective
or ineffective or inconclusive, and closed, in that order; an
ineffective or inconclusive action closes only as Step 4.4 allows, else
it is reopened, then superseded by a replacement at proposed.
Let no prose, merge, or local green skip a state.
Gotchas
- An escape-path audit taken from today's list of gates describes the
battery after the fix landed, so every gate looks fine.
- A
Risk-reduction claim of "prevent" with no baseline, target, or horizon
can never be marked effective or ineffective at the review date, only
inconclusive, permanently.
Common mistakes
- “Human error” or “reviewer missed it” instead of the conditions and the
absent gate that let the action escape.
- Publishing sensitive incident artifacts without access and retention
controls.
1---2name: post-mortem3description: "Explains why safeguards missed a production escape, incident, loss, or failed work cycle, or why its impact grew, and what prevents it happening again. Use when a production escape, incident, material loss, or failed work cycle has been contained, or when the user asks how this reached production or what prevents another failed handoff. Establish any technical cause and immediate containment first; a process-only failure needs its event evidence. Skip ordinary bugs and ongoing uncontained impact."4metadata:5 source: https://skillmd.com/skills/augments-labs/post-mortem6---78# Post-Mortem910Explain why the failure escaped or grew, then measure whether each corrective11control reduces recurrence, detection, or impact risk. Take the code-level12cause from `debugging`; do not work it out again. Close on controls that are13deployed and falsified, never on reflection. For a process-only failure, use14the observed trail of events and decisions; do not invent a code defect to15enter this skill.1617## When to use1819- A failure reached users or production, escaped far downstream, caused real20 loss, or exposed a process failure worth correcting.21- **Skip** an ordinary reproduced bug. For a technical incident, establish the22 technical cause and immediate containment first: invoke `debugging` if the23 cause is unknown, `containing-an-incident` if impact is ongoing.2425## Step 1: Control the record26271. Open `assets/post-mortem-template.md` when starting the analysis. Fill28 `Control and evidence handling` first.292. Fill `Summary and impact` and `Timeline` from artifacts (any containment30 record first), not from memory. Mark every time `observed` or `estimated`.3132## Step 2: Find the structural cause33341. Fill `Root cause and contributing conditions`: cite the technical cause35 from `debugging` when applicable, or the evidence of the process failure,36 then each condition that made the introduction or the impact more likely.37 Name conditions, never a person. When there are several conditions, keep38 them separate.392. Fill `Escape-path audit`: freeze the inventory of gates and surfaces live40 when the failure escaped, with the hash of its source, then fill every41 column for each entry, and for each omission record a decision that42 someone is accountable for.433. Fill every bullet of `Risk-reduction claim` for each action. Never claim44 that recurrence is impossible.454. Fill the `Corrective actions` rows, each mapped to a structural cause;46 receipts and `External state` stay `pending` and live in the ledger.475. Challenge the draft. Fill `assets/escape-path-challenger.md` before48 dispatching it read-only as `dispatching-parallel-agents` Step 2 says.49 Without a challenger, record the omission the `Expected reviewers` field50 requires, approved if the failure reached users or caused loss; a51 self-review never stands in. Record a decision on every finding. A52 `findings` verdict blocks issuing the analysis until the corrected draft53 is challenged again, twice at most; findings still open then go to the54 approver.556. Issue the analysis, which is never edited after issue, with its56 `External lifecycle ledger` (the ledger: add to it, never rewrite it), to57 the user-set path or the template's path, under the storage authority held58 now, and record its identity as the template says. If the record is in the59 repository, it is a new state under check.6061## Step 3: Propose and get the actions accepted62631. Record each corrective action's `proposed` state in the ledger.642. Present:6566 ```text67 Post-mortem {{identity}} — {{path}}68 Impact: {{one line}} Structural cause: {{one line}} Escaped gate: {{one line}}69 Actions: {{each with owner and date}}7071 1. Accept the actions72 2. Request changes73 3. Reject the analysis74 4. Cancel7576 Recommendation: {{the answer the evidence and ownership support}} — {{one sentence}}.77 ```7879 Ask this and end the turn; `clarifying-intent` owns what closes it.803. Move an action out of `proposed` only on a complete trusted receipt (proof,81 from a trusted source outside the state under check, that an action or82 decision really happened) for the exact scope and dates. Option 2 issues83 a successor analysis and asks again; 3 and 4 record each action rejected84 or cancelled, with its residual risk.8586## Step 4: Prove and track the controls87881. Falsify every corrective gate you implement. Record the targeted89 fail-then-pass proof in the ledger as raw evidence, never in the issued90 analysis: it fails on the captured incident or a representative bad91 case, it passes on the good control, and the version before the fix did92 not catch it.932. Track each control to a receipt bound to an identity, from the surface94 that enforces it: CI, runtime, review, release, alerting, recovery. Never95 infer from a summary.963. **REQUIRED SUB-SKILL:** invoke `verification-strategy` for any change to97 the battery itself (the set of checks the project runs). Add no duplicate98 check nobody owns.994. At the date declared in advance, compare the baseline against the target100 and confirm that the gate ran. Write `effective`, `ineffective`, or101 `inconclusive`. If it is `ineffective` or `inconclusive`, reopen, unless102 the exact approver rule accepts closure with the residual risk.103104## Action states105106Record in the ledger, outside the issued analysis, each action's107move from `proposed` to one of `{owner-accepted | rejected | cancelled |108superseded}`, and, for an accepted action, from there through `implemented`,109`falsified`, `deployed/enforced`, `effectiveness reviewed`, then `effective`110or `ineffective` or `inconclusive`, and `closed`, in that order; an111`ineffective` or `inconclusive` action closes only as Step 4.4 allows, else112it is `reopened`, then `superseded` by a replacement at `proposed`.113Let no prose, merge, or local green skip a state.114115## Gotchas116117- An escape-path audit taken from today's list of gates describes the118 battery after the fix landed, so every gate looks fine.119- A `Risk-reduction claim` of "prevent" with no baseline, target, or horizon120 can never be marked `effective` or `ineffective` at the review date, only121 `inconclusive`, permanently.122123## Common mistakes124125- “Human error” or “reviewer missed it” instead of the conditions and the126 absent gate that let the action escape.127- Publishing sensitive incident artifacts without access and retention128 controls.